Replacing the PM with the Full-Stack Builder: How LinkedIn Is Rebuilding Product Development for AI

Cohen frames the case for change as both an opportunity and a necessity. He says we are entering a phase where the time constant of change is far greater than the time constant of response, meaning change happens faster than organizations can react. LinkedIn's data on the world of work gives him evidence: by 2030, the skills required for a given job will change by 70 percent, and the fastest-growing jobs are growing at more than 70 percent over last year's fastest-growing ones.

Against that backdrop, he looks at the builder, someone who takes an idea and brings it to life. The classic steps of researching, specifying, designing, coding, launching and iterating each grew into many sub-steps. Research alone now draws on 10 to 15 information sources, and design, privacy and security reviews each add their own gates. Every sub-step has a valid reason to exist, but together they mean a small feature needs multiple teams, code bases and sprints.

Cohen says process complexity then turned into organizational complexity and micro-specialization, with interaction, animation and content design as separate crafts on top of the engineering, product and design split. He shows a diagram combining both layers that tends to stun people who only see their own corner. The opportunity, as he sees it, is to collapse the stack and return to craftsmanship by rethinking the whole product development life cycle.

The full-stack builder model aims to let great builders take an idea to market regardless of role or team, combining skills that used to belong to separate domains. Cohen stresses that it is not a sequence of steps but a fluid interaction between human and machine. He wants people to spend their time on vision, empathy, communication, creativity and judgment, while he works "really, really hard" to automate everything else.

He singles out judgment, the ability to make high-quality decisions in complex, ambiguous situations, as the most important trait. On creativity, he says AI is not yet great at the next-level kind that goes beyond the obvious. The payoff he describes is not just more shots on goal but an organization that is more nimble, adaptive and resilient, able to match the pace of response to the pace of change.

His analogy is Navy SEALs, who are cross-trained across areas, specialize in the mission and work in small pods that can be assembled quickly. He clarifies that builders don't work alone and that he still believes in teams, only smaller and more focused ones. At LinkedIn, pods built around full-stack builders tackle a problem for roughly a quarter and are then reassembled, with early gains in velocity and focus.

Cohen says startups built AI-natively can simply start as full-stack builders, but a company at LinkedIn's scale needs three components: platform, tools and agents, and culture. The platform work means re-architecting core systems, including composable UI components with server-side support, so AI can reason over them. He says a third-party tool never works on the LinkedIn stack out of the box, so the company works almost in alpha mode with vendors, including coding and design tools, to adapt them.

On the agent side, each head of craft builds their own. A trust agent can walk a spec through potential harm vectors; when LinkedIn ran an older "open to work" spec through it, it found the issues the team had foreseen as well as holes they caught only later. A growth agent, trained on LinkedIn's loops, funnels and past tests, critiques how good an idea is, and a research agent trained on member personas critiqued a marketing spec well enough to shift a team's direction.

There is also an analyst agent that queries LinkedIn's enormous graph without SQL or a data science team, and Cohen calls all of them "MVP plus plus." The team uses Copilot Enterprise and ChatGPT Enterprise for knowledge agents, but says its own customization, including an orchestrator so agents work with one another, produced the biggest gains. Different teams also gravitate to different design tools such as Figma, Subframe and Magic Patterns, so convergence is a challenge, and he doubts there will be a winner-takes-all.

Cohen splits the pipeline into idea-to-design and code-to-launch. Coding got attention first, with a coding agent and a maintenance agent that handles failed builds; he says close to half of those builds are now handled by the maintenance agent and a QA agent. The idea-to-design stage, where much of a product's quality originates, had little investment until this program began.

The first agent MVPs arrived about four to five months after the program was announced internally at the end of last year, though the dedicated work took only a couple of months. Most of that work was gathering and cleaning the corpus. Cohen says pointing a model at an entire drive does a poor job, because it can't judge the importance of past material, so the team curated "golden examples" and decided what context each agent should focus on.

He ties this to more than a decade ago, when he helped rebuild the LinkedIn feed and spent weeks sorting good professional posts from bad. The lesson was the same: the right data matters more than all the data. His advice for anyone learning AI tools is to be ready to invest hours up front before velocity and quality improve.

Cohen describes the benefit as experimentation volume multiplied by quality, divided by time from idea to launch. A substantial part of the organization already uses the tools, though not yet a large share, and PMs, designers and engineers are saving hours a week. Designers and PMs are picking up bugs directly from Jira tickets, and the biggest current demand is that everybody wants access.

A core team runs the full-stack builder track across R&D, with pods that agree to give feedback in return for access. The team behind LinkedIn's semantic people and job search used the tools, with PMs building their own dashboards, and designers have begun pushing pull requests. At the junior end, the associate product manager program ends this year and an associate product builder program starts in January, teaching people to code, design and do product work at LinkedIn.

Lenny observes that top performers seem to benefit most, and Cohen agrees: top talent keeps trying to get better at its craft and wants to be at the cutting edge. He is clear, however, that giving people tools isn't enough. Looking back at LinkedIn's shift from desktop to mobile, he says adoption needs incentives, motivation and examples of peers succeeding.

Cohen says only about the cutting-edge 5 percent of talent adopts new tools on its own, so the vast majority needs real change management. Platforms and tools are a prerequisite but not sufficient. The cultural levers include defining new expectations in hiring, calibration and evaluation, building a sense of agency so people want to improve the tools, piloting success in visible pods, and celebrating wins at all-hands meetings.

He shares examples: a partnerships leader built a developer portal connector himself instead of waiting for an engineer, and a user researcher moved into an open growth PM role after using the new tools to show she could do it. Cohen says top-down directives aren't enough, because people need to feel that success stories exist and the movement is worth their time. LinkedIn is also folding the approach into 360 reviews and the upcoming bi-annual performance cycle.

He is deliberately not rolling out the title everywhere quickly, saying that would dilute it, and that he cares about mindset more than titles. His message is that anyone waiting for a re-org or declaration to start working differently is waiting too long: use whatever tools exist, build one if needed, and show examples. He also encourages people to share tools they find useful, with a focus on finding one thing that works and investing in it.

Asked about surprises, Cohen repeats that off-the-shelf tools never worked, that giving access to all context failed badly, and that convergence on tools proved hard. Quality has been better, though he cautions that this is still early adopters iterating. He also says some people don't want to be full-stack builders, which is fine, since specialists and system builders who empower full-stack builders still have a role, just not as many as before.

LinkedIn now has a formal full-stack builder title and a career ladder forming around it. Cohen suggests leaders look across their organizations to see who already flexes across engineering, design, product and even business development. He thinks design may be the hardest craft for agents to learn, giving designers a slight edge, but says it is mostly mindset, agency and a growth orientation.

He sees this as the easiest time ever to move between product roles, and he tells the team that personal and organizational incentives are aligned: LinkedIn needs more adaptive people, and people want to be at the cutting edge for their own careers.

Cohen's advice for others is to think about the platform, then the tools, then spend most of the time on culture. One thing he would do differently is to share more with the wider organization earlier, because while his core team worked closely on the program, others kept asking who was doing what. Starting small is fine, but visibility across the organization matters.

He contrasts a startup that can build something in a week with transforming a large company, which requires high ambition and impatience about the goal but patience about how to get there. Skipping platform investment or customization, he says, leaves you with vanilla generic agents, and expecting 2x productivity in a week is unrealistic. Collect examples and think about the transformation.

His closing advice is not to wait. He announced the new mode upfront, started in pockets and pods, and framed it as continuous progress rather than an end state since the way of building will transform every few years. Leaders should over-communicate progress and even set KPIs or OKRs for themselves, and employees whose leaders aren't announcing a program should go do it or join an organization that does.

In the lightning round, Cohen recommends Why Nations Fail, which contrasts extractive and inclusive institutions, Peter Attia's Outlive, and David Deutsch's The Beginning of Infinity, which he values for its emphasis on finding good explanations for why things happen. He also likes a Hebrew podcast called One Song, and his product wish is a steering-wheel button that summons an AI friend in the passenger seat, similar to ChatGPT voice mode, which Lenny notes Teslas now do with Grok.

His motto comes from growth-mindset thinking: "becoming is better than being." He ties it to the builder mindset of continuous iteration and asks himself what the delta is between his 2026 self and his 2025 self. He then reveals that he is leaving LinkedIn after 14 years, saying he is proud, grateful for the tough situations that taught him the most, and excited by his bias for change, and he will always think of LinkedIn as one of the babies he helped grow.